xref: /openbmc/linux/fs/btrfs/delayed-ref.c (revision 330e261c)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2009 Oracle.  All rights reserved.
4  */
5 
6 #include <linux/sched.h>
7 #include <linux/slab.h>
8 #include <linux/sort.h>
9 #include "ctree.h"
10 #include "delayed-ref.h"
11 #include "transaction.h"
12 #include "qgroup.h"
13 
14 struct kmem_cache *btrfs_delayed_ref_head_cachep;
15 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
16 struct kmem_cache *btrfs_delayed_data_ref_cachep;
17 struct kmem_cache *btrfs_delayed_extent_op_cachep;
18 /*
19  * delayed back reference update tracking.  For subvolume trees
20  * we queue up extent allocations and backref maintenance for
21  * delayed processing.   This avoids deep call chains where we
22  * add extents in the middle of btrfs_search_slot, and it allows
23  * us to buffer up frequently modified backrefs in an rb tree instead
24  * of hammering updates on the extent allocation tree.
25  */
26 
27 /*
28  * compare two delayed tree backrefs with same bytenr and type
29  */
30 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
31 			  struct btrfs_delayed_tree_ref *ref2)
32 {
33 	if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
34 		if (ref1->root < ref2->root)
35 			return -1;
36 		if (ref1->root > ref2->root)
37 			return 1;
38 	} else {
39 		if (ref1->parent < ref2->parent)
40 			return -1;
41 		if (ref1->parent > ref2->parent)
42 			return 1;
43 	}
44 	return 0;
45 }
46 
47 /*
48  * compare two delayed data backrefs with same bytenr and type
49  */
50 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
51 			  struct btrfs_delayed_data_ref *ref2)
52 {
53 	if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
54 		if (ref1->root < ref2->root)
55 			return -1;
56 		if (ref1->root > ref2->root)
57 			return 1;
58 		if (ref1->objectid < ref2->objectid)
59 			return -1;
60 		if (ref1->objectid > ref2->objectid)
61 			return 1;
62 		if (ref1->offset < ref2->offset)
63 			return -1;
64 		if (ref1->offset > ref2->offset)
65 			return 1;
66 	} else {
67 		if (ref1->parent < ref2->parent)
68 			return -1;
69 		if (ref1->parent > ref2->parent)
70 			return 1;
71 	}
72 	return 0;
73 }
74 
75 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
76 		     struct btrfs_delayed_ref_node *ref2,
77 		     bool check_seq)
78 {
79 	int ret = 0;
80 
81 	if (ref1->type < ref2->type)
82 		return -1;
83 	if (ref1->type > ref2->type)
84 		return 1;
85 	if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
86 	    ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
87 		ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
88 				     btrfs_delayed_node_to_tree_ref(ref2));
89 	else
90 		ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
91 				     btrfs_delayed_node_to_data_ref(ref2));
92 	if (ret)
93 		return ret;
94 	if (check_seq) {
95 		if (ref1->seq < ref2->seq)
96 			return -1;
97 		if (ref1->seq > ref2->seq)
98 			return 1;
99 	}
100 	return 0;
101 }
102 
103 /* insert a new ref to head ref rbtree */
104 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root *root,
105 						   struct rb_node *node)
106 {
107 	struct rb_node **p = &root->rb_node;
108 	struct rb_node *parent_node = NULL;
109 	struct btrfs_delayed_ref_head *entry;
110 	struct btrfs_delayed_ref_head *ins;
111 	u64 bytenr;
112 
113 	ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
114 	bytenr = ins->bytenr;
115 	while (*p) {
116 		parent_node = *p;
117 		entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
118 				 href_node);
119 
120 		if (bytenr < entry->bytenr)
121 			p = &(*p)->rb_left;
122 		else if (bytenr > entry->bytenr)
123 			p = &(*p)->rb_right;
124 		else
125 			return entry;
126 	}
127 
128 	rb_link_node(node, parent_node, p);
129 	rb_insert_color(node, root);
130 	return NULL;
131 }
132 
133 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root *root,
134 		struct btrfs_delayed_ref_node *ins)
135 {
136 	struct rb_node **p = &root->rb_node;
137 	struct rb_node *node = &ins->ref_node;
138 	struct rb_node *parent_node = NULL;
139 	struct btrfs_delayed_ref_node *entry;
140 
141 	while (*p) {
142 		int comp;
143 
144 		parent_node = *p;
145 		entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
146 				 ref_node);
147 		comp = comp_refs(ins, entry, true);
148 		if (comp < 0)
149 			p = &(*p)->rb_left;
150 		else if (comp > 0)
151 			p = &(*p)->rb_right;
152 		else
153 			return entry;
154 	}
155 
156 	rb_link_node(node, parent_node, p);
157 	rb_insert_color(node, root);
158 	return NULL;
159 }
160 
161 /*
162  * find an head entry based on bytenr. This returns the delayed ref
163  * head if it was able to find one, or NULL if nothing was in that spot.
164  * If return_bigger is given, the next bigger entry is returned if no exact
165  * match is found.
166  */
167 static struct btrfs_delayed_ref_head *
168 find_ref_head(struct rb_root *root, u64 bytenr,
169 	      int return_bigger)
170 {
171 	struct rb_node *n;
172 	struct btrfs_delayed_ref_head *entry;
173 
174 	n = root->rb_node;
175 	entry = NULL;
176 	while (n) {
177 		entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
178 
179 		if (bytenr < entry->bytenr)
180 			n = n->rb_left;
181 		else if (bytenr > entry->bytenr)
182 			n = n->rb_right;
183 		else
184 			return entry;
185 	}
186 	if (entry && return_bigger) {
187 		if (bytenr > entry->bytenr) {
188 			n = rb_next(&entry->href_node);
189 			if (!n)
190 				n = rb_first(root);
191 			entry = rb_entry(n, struct btrfs_delayed_ref_head,
192 					 href_node);
193 			return entry;
194 		}
195 		return entry;
196 	}
197 	return NULL;
198 }
199 
200 int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
201 			   struct btrfs_delayed_ref_head *head)
202 {
203 	struct btrfs_delayed_ref_root *delayed_refs;
204 
205 	delayed_refs = &trans->transaction->delayed_refs;
206 	lockdep_assert_held(&delayed_refs->lock);
207 	if (mutex_trylock(&head->mutex))
208 		return 0;
209 
210 	refcount_inc(&head->refs);
211 	spin_unlock(&delayed_refs->lock);
212 
213 	mutex_lock(&head->mutex);
214 	spin_lock(&delayed_refs->lock);
215 	if (RB_EMPTY_NODE(&head->href_node)) {
216 		mutex_unlock(&head->mutex);
217 		btrfs_put_delayed_ref_head(head);
218 		return -EAGAIN;
219 	}
220 	btrfs_put_delayed_ref_head(head);
221 	return 0;
222 }
223 
224 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
225 				    struct btrfs_delayed_ref_root *delayed_refs,
226 				    struct btrfs_delayed_ref_head *head,
227 				    struct btrfs_delayed_ref_node *ref)
228 {
229 	lockdep_assert_held(&head->lock);
230 	rb_erase(&ref->ref_node, &head->ref_tree);
231 	RB_CLEAR_NODE(&ref->ref_node);
232 	if (!list_empty(&ref->add_list))
233 		list_del(&ref->add_list);
234 	ref->in_tree = 0;
235 	btrfs_put_delayed_ref(ref);
236 	atomic_dec(&delayed_refs->num_entries);
237 	if (trans->delayed_ref_updates)
238 		trans->delayed_ref_updates--;
239 }
240 
241 static bool merge_ref(struct btrfs_trans_handle *trans,
242 		      struct btrfs_delayed_ref_root *delayed_refs,
243 		      struct btrfs_delayed_ref_head *head,
244 		      struct btrfs_delayed_ref_node *ref,
245 		      u64 seq)
246 {
247 	struct btrfs_delayed_ref_node *next;
248 	struct rb_node *node = rb_next(&ref->ref_node);
249 	bool done = false;
250 
251 	while (!done && node) {
252 		int mod;
253 
254 		next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
255 		node = rb_next(node);
256 		if (seq && next->seq >= seq)
257 			break;
258 		if (comp_refs(ref, next, false))
259 			break;
260 
261 		if (ref->action == next->action) {
262 			mod = next->ref_mod;
263 		} else {
264 			if (ref->ref_mod < next->ref_mod) {
265 				swap(ref, next);
266 				done = true;
267 			}
268 			mod = -next->ref_mod;
269 		}
270 
271 		drop_delayed_ref(trans, delayed_refs, head, next);
272 		ref->ref_mod += mod;
273 		if (ref->ref_mod == 0) {
274 			drop_delayed_ref(trans, delayed_refs, head, ref);
275 			done = true;
276 		} else {
277 			/*
278 			 * Can't have multiples of the same ref on a tree block.
279 			 */
280 			WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
281 				ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
282 		}
283 	}
284 
285 	return done;
286 }
287 
288 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
289 			      struct btrfs_fs_info *fs_info,
290 			      struct btrfs_delayed_ref_root *delayed_refs,
291 			      struct btrfs_delayed_ref_head *head)
292 {
293 	struct btrfs_delayed_ref_node *ref;
294 	struct rb_node *node;
295 	u64 seq = 0;
296 
297 	lockdep_assert_held(&head->lock);
298 
299 	if (RB_EMPTY_ROOT(&head->ref_tree))
300 		return;
301 
302 	/* We don't have too many refs to merge for data. */
303 	if (head->is_data)
304 		return;
305 
306 	spin_lock(&fs_info->tree_mod_seq_lock);
307 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
308 		struct seq_list *elem;
309 
310 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
311 					struct seq_list, list);
312 		seq = elem->seq;
313 	}
314 	spin_unlock(&fs_info->tree_mod_seq_lock);
315 
316 again:
317 	for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
318 		ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
319 		if (seq && ref->seq >= seq)
320 			continue;
321 		if (merge_ref(trans, delayed_refs, head, ref, seq))
322 			goto again;
323 	}
324 }
325 
326 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
327 			    struct btrfs_delayed_ref_root *delayed_refs,
328 			    u64 seq)
329 {
330 	struct seq_list *elem;
331 	int ret = 0;
332 
333 	spin_lock(&fs_info->tree_mod_seq_lock);
334 	if (!list_empty(&fs_info->tree_mod_seq_list)) {
335 		elem = list_first_entry(&fs_info->tree_mod_seq_list,
336 					struct seq_list, list);
337 		if (seq >= elem->seq) {
338 			btrfs_debug(fs_info,
339 				"holding back delayed_ref %#x.%x, lowest is %#x.%x (%p)",
340 				(u32)(seq >> 32), (u32)seq,
341 				(u32)(elem->seq >> 32), (u32)elem->seq,
342 				delayed_refs);
343 			ret = 1;
344 		}
345 	}
346 
347 	spin_unlock(&fs_info->tree_mod_seq_lock);
348 	return ret;
349 }
350 
351 struct btrfs_delayed_ref_head *
352 btrfs_select_ref_head(struct btrfs_trans_handle *trans)
353 {
354 	struct btrfs_delayed_ref_root *delayed_refs;
355 	struct btrfs_delayed_ref_head *head;
356 	u64 start;
357 	bool loop = false;
358 
359 	delayed_refs = &trans->transaction->delayed_refs;
360 
361 again:
362 	start = delayed_refs->run_delayed_start;
363 	head = find_ref_head(&delayed_refs->href_root, start, 1);
364 	if (!head && !loop) {
365 		delayed_refs->run_delayed_start = 0;
366 		start = 0;
367 		loop = true;
368 		head = find_ref_head(&delayed_refs->href_root, start, 1);
369 		if (!head)
370 			return NULL;
371 	} else if (!head && loop) {
372 		return NULL;
373 	}
374 
375 	while (head->processing) {
376 		struct rb_node *node;
377 
378 		node = rb_next(&head->href_node);
379 		if (!node) {
380 			if (loop)
381 				return NULL;
382 			delayed_refs->run_delayed_start = 0;
383 			start = 0;
384 			loop = true;
385 			goto again;
386 		}
387 		head = rb_entry(node, struct btrfs_delayed_ref_head,
388 				href_node);
389 	}
390 
391 	head->processing = 1;
392 	WARN_ON(delayed_refs->num_heads_ready == 0);
393 	delayed_refs->num_heads_ready--;
394 	delayed_refs->run_delayed_start = head->bytenr +
395 		head->num_bytes;
396 	return head;
397 }
398 
399 /*
400  * Helper to insert the ref_node to the tail or merge with tail.
401  *
402  * Return 0 for insert.
403  * Return >0 for merge.
404  */
405 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
406 			      struct btrfs_delayed_ref_root *root,
407 			      struct btrfs_delayed_ref_head *href,
408 			      struct btrfs_delayed_ref_node *ref)
409 {
410 	struct btrfs_delayed_ref_node *exist;
411 	int mod;
412 	int ret = 0;
413 
414 	spin_lock(&href->lock);
415 	exist = tree_insert(&href->ref_tree, ref);
416 	if (!exist)
417 		goto inserted;
418 
419 	/* Now we are sure we can merge */
420 	ret = 1;
421 	if (exist->action == ref->action) {
422 		mod = ref->ref_mod;
423 	} else {
424 		/* Need to change action */
425 		if (exist->ref_mod < ref->ref_mod) {
426 			exist->action = ref->action;
427 			mod = -exist->ref_mod;
428 			exist->ref_mod = ref->ref_mod;
429 			if (ref->action == BTRFS_ADD_DELAYED_REF)
430 				list_add_tail(&exist->add_list,
431 					      &href->ref_add_list);
432 			else if (ref->action == BTRFS_DROP_DELAYED_REF) {
433 				ASSERT(!list_empty(&exist->add_list));
434 				list_del(&exist->add_list);
435 			} else {
436 				ASSERT(0);
437 			}
438 		} else
439 			mod = -ref->ref_mod;
440 	}
441 	exist->ref_mod += mod;
442 
443 	/* remove existing tail if its ref_mod is zero */
444 	if (exist->ref_mod == 0)
445 		drop_delayed_ref(trans, root, href, exist);
446 	spin_unlock(&href->lock);
447 	return ret;
448 inserted:
449 	if (ref->action == BTRFS_ADD_DELAYED_REF)
450 		list_add_tail(&ref->add_list, &href->ref_add_list);
451 	atomic_inc(&root->num_entries);
452 	trans->delayed_ref_updates++;
453 	spin_unlock(&href->lock);
454 	return ret;
455 }
456 
457 /*
458  * helper function to update the accounting in the head ref
459  * existing and update must have the same bytenr
460  */
461 static noinline void
462 update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
463 			 struct btrfs_delayed_ref_head *existing,
464 			 struct btrfs_delayed_ref_head *update,
465 			 int *old_ref_mod_ret)
466 {
467 	int old_ref_mod;
468 
469 	BUG_ON(existing->is_data != update->is_data);
470 
471 	spin_lock(&existing->lock);
472 	if (update->must_insert_reserved) {
473 		/* if the extent was freed and then
474 		 * reallocated before the delayed ref
475 		 * entries were processed, we can end up
476 		 * with an existing head ref without
477 		 * the must_insert_reserved flag set.
478 		 * Set it again here
479 		 */
480 		existing->must_insert_reserved = update->must_insert_reserved;
481 
482 		/*
483 		 * update the num_bytes so we make sure the accounting
484 		 * is done correctly
485 		 */
486 		existing->num_bytes = update->num_bytes;
487 
488 	}
489 
490 	if (update->extent_op) {
491 		if (!existing->extent_op) {
492 			existing->extent_op = update->extent_op;
493 		} else {
494 			if (update->extent_op->update_key) {
495 				memcpy(&existing->extent_op->key,
496 				       &update->extent_op->key,
497 				       sizeof(update->extent_op->key));
498 				existing->extent_op->update_key = true;
499 			}
500 			if (update->extent_op->update_flags) {
501 				existing->extent_op->flags_to_set |=
502 					update->extent_op->flags_to_set;
503 				existing->extent_op->update_flags = true;
504 			}
505 			btrfs_free_delayed_extent_op(update->extent_op);
506 		}
507 	}
508 	/*
509 	 * update the reference mod on the head to reflect this new operation,
510 	 * only need the lock for this case cause we could be processing it
511 	 * currently, for refs we just added we know we're a-ok.
512 	 */
513 	old_ref_mod = existing->total_ref_mod;
514 	if (old_ref_mod_ret)
515 		*old_ref_mod_ret = old_ref_mod;
516 	existing->ref_mod += update->ref_mod;
517 	existing->total_ref_mod += update->ref_mod;
518 
519 	/*
520 	 * If we are going to from a positive ref mod to a negative or vice
521 	 * versa we need to make sure to adjust pending_csums accordingly.
522 	 */
523 	if (existing->is_data) {
524 		if (existing->total_ref_mod >= 0 && old_ref_mod < 0)
525 			delayed_refs->pending_csums -= existing->num_bytes;
526 		if (existing->total_ref_mod < 0 && old_ref_mod >= 0)
527 			delayed_refs->pending_csums += existing->num_bytes;
528 	}
529 	spin_unlock(&existing->lock);
530 }
531 
532 /*
533  * helper function to actually insert a head node into the rbtree.
534  * this does all the dirty work in terms of maintaining the correct
535  * overall modification count.
536  */
537 static noinline struct btrfs_delayed_ref_head *
538 add_delayed_ref_head(struct btrfs_fs_info *fs_info,
539 		     struct btrfs_trans_handle *trans,
540 		     struct btrfs_delayed_ref_head *head_ref,
541 		     struct btrfs_qgroup_extent_record *qrecord,
542 		     u64 bytenr, u64 num_bytes, u64 ref_root, u64 reserved,
543 		     int action, int is_data, int is_system,
544 		     int *qrecord_inserted_ret,
545 		     int *old_ref_mod, int *new_ref_mod)
546 
547 {
548 	struct btrfs_delayed_ref_head *existing;
549 	struct btrfs_delayed_ref_root *delayed_refs;
550 	int count_mod = 1;
551 	int must_insert_reserved = 0;
552 	int qrecord_inserted = 0;
553 
554 	/* If reserved is provided, it must be a data extent. */
555 	BUG_ON(!is_data && reserved);
556 
557 	/*
558 	 * the head node stores the sum of all the mods, so dropping a ref
559 	 * should drop the sum in the head node by one.
560 	 */
561 	if (action == BTRFS_UPDATE_DELAYED_HEAD)
562 		count_mod = 0;
563 	else if (action == BTRFS_DROP_DELAYED_REF)
564 		count_mod = -1;
565 
566 	/*
567 	 * BTRFS_ADD_DELAYED_EXTENT means that we need to update
568 	 * the reserved accounting when the extent is finally added, or
569 	 * if a later modification deletes the delayed ref without ever
570 	 * inserting the extent into the extent allocation tree.
571 	 * ref->must_insert_reserved is the flag used to record
572 	 * that accounting mods are required.
573 	 *
574 	 * Once we record must_insert_reserved, switch the action to
575 	 * BTRFS_ADD_DELAYED_REF because other special casing is not required.
576 	 */
577 	if (action == BTRFS_ADD_DELAYED_EXTENT)
578 		must_insert_reserved = 1;
579 	else
580 		must_insert_reserved = 0;
581 
582 	delayed_refs = &trans->transaction->delayed_refs;
583 
584 	refcount_set(&head_ref->refs, 1);
585 	head_ref->bytenr = bytenr;
586 	head_ref->num_bytes = num_bytes;
587 	head_ref->ref_mod = count_mod;
588 	head_ref->must_insert_reserved = must_insert_reserved;
589 	head_ref->is_data = is_data;
590 	head_ref->is_system = is_system;
591 	head_ref->ref_tree = RB_ROOT;
592 	INIT_LIST_HEAD(&head_ref->ref_add_list);
593 	RB_CLEAR_NODE(&head_ref->href_node);
594 	head_ref->processing = 0;
595 	head_ref->total_ref_mod = count_mod;
596 	head_ref->qgroup_reserved = 0;
597 	head_ref->qgroup_ref_root = 0;
598 	spin_lock_init(&head_ref->lock);
599 	mutex_init(&head_ref->mutex);
600 
601 	/* Record qgroup extent info if provided */
602 	if (qrecord) {
603 		if (ref_root && reserved) {
604 			head_ref->qgroup_ref_root = ref_root;
605 			head_ref->qgroup_reserved = reserved;
606 		}
607 
608 		qrecord->bytenr = bytenr;
609 		qrecord->num_bytes = num_bytes;
610 		qrecord->old_roots = NULL;
611 
612 		if(btrfs_qgroup_trace_extent_nolock(fs_info,
613 					delayed_refs, qrecord))
614 			kfree(qrecord);
615 		else
616 			qrecord_inserted = 1;
617 	}
618 
619 	trace_add_delayed_ref_head(fs_info, head_ref, action);
620 
621 	existing = htree_insert(&delayed_refs->href_root,
622 				&head_ref->href_node);
623 	if (existing) {
624 		WARN_ON(ref_root && reserved && existing->qgroup_ref_root
625 			&& existing->qgroup_reserved);
626 		update_existing_head_ref(delayed_refs, existing, head_ref,
627 					 old_ref_mod);
628 		/*
629 		 * we've updated the existing ref, free the newly
630 		 * allocated ref
631 		 */
632 		kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
633 		head_ref = existing;
634 	} else {
635 		if (old_ref_mod)
636 			*old_ref_mod = 0;
637 		if (is_data && count_mod < 0)
638 			delayed_refs->pending_csums += num_bytes;
639 		delayed_refs->num_heads++;
640 		delayed_refs->num_heads_ready++;
641 		atomic_inc(&delayed_refs->num_entries);
642 		trans->delayed_ref_updates++;
643 	}
644 	if (qrecord_inserted_ret)
645 		*qrecord_inserted_ret = qrecord_inserted;
646 	if (new_ref_mod)
647 		*new_ref_mod = head_ref->total_ref_mod;
648 	return head_ref;
649 }
650 
651 /*
652  * helper to insert a delayed tree ref into the rbtree.
653  */
654 static noinline void
655 add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
656 		     struct btrfs_trans_handle *trans,
657 		     struct btrfs_delayed_ref_head *head_ref,
658 		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
659 		     u64 num_bytes, u64 parent, u64 ref_root, int level,
660 		     int action)
661 {
662 	struct btrfs_delayed_tree_ref *full_ref;
663 	struct btrfs_delayed_ref_root *delayed_refs;
664 	u64 seq = 0;
665 	int ret;
666 
667 	if (action == BTRFS_ADD_DELAYED_EXTENT)
668 		action = BTRFS_ADD_DELAYED_REF;
669 
670 	if (is_fstree(ref_root))
671 		seq = atomic64_read(&fs_info->tree_mod_seq);
672 	delayed_refs = &trans->transaction->delayed_refs;
673 
674 	/* first set the basic ref node struct up */
675 	refcount_set(&ref->refs, 1);
676 	ref->bytenr = bytenr;
677 	ref->num_bytes = num_bytes;
678 	ref->ref_mod = 1;
679 	ref->action = action;
680 	ref->is_head = 0;
681 	ref->in_tree = 1;
682 	ref->seq = seq;
683 	RB_CLEAR_NODE(&ref->ref_node);
684 	INIT_LIST_HEAD(&ref->add_list);
685 
686 	full_ref = btrfs_delayed_node_to_tree_ref(ref);
687 	full_ref->parent = parent;
688 	full_ref->root = ref_root;
689 	if (parent)
690 		ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
691 	else
692 		ref->type = BTRFS_TREE_BLOCK_REF_KEY;
693 	full_ref->level = level;
694 
695 	trace_add_delayed_tree_ref(fs_info, ref, full_ref, action);
696 
697 	ret = insert_delayed_ref(trans, delayed_refs, head_ref, ref);
698 
699 	/*
700 	 * XXX: memory should be freed at the same level allocated.
701 	 * But bad practice is anywhere... Follow it now. Need cleanup.
702 	 */
703 	if (ret > 0)
704 		kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
705 }
706 
707 /*
708  * helper to insert a delayed data ref into the rbtree.
709  */
710 static noinline void
711 add_delayed_data_ref(struct btrfs_fs_info *fs_info,
712 		     struct btrfs_trans_handle *trans,
713 		     struct btrfs_delayed_ref_head *head_ref,
714 		     struct btrfs_delayed_ref_node *ref, u64 bytenr,
715 		     u64 num_bytes, u64 parent, u64 ref_root, u64 owner,
716 		     u64 offset, int action)
717 {
718 	struct btrfs_delayed_data_ref *full_ref;
719 	struct btrfs_delayed_ref_root *delayed_refs;
720 	u64 seq = 0;
721 	int ret;
722 
723 	if (action == BTRFS_ADD_DELAYED_EXTENT)
724 		action = BTRFS_ADD_DELAYED_REF;
725 
726 	delayed_refs = &trans->transaction->delayed_refs;
727 
728 	if (is_fstree(ref_root))
729 		seq = atomic64_read(&fs_info->tree_mod_seq);
730 
731 	/* first set the basic ref node struct up */
732 	refcount_set(&ref->refs, 1);
733 	ref->bytenr = bytenr;
734 	ref->num_bytes = num_bytes;
735 	ref->ref_mod = 1;
736 	ref->action = action;
737 	ref->is_head = 0;
738 	ref->in_tree = 1;
739 	ref->seq = seq;
740 	RB_CLEAR_NODE(&ref->ref_node);
741 	INIT_LIST_HEAD(&ref->add_list);
742 
743 	full_ref = btrfs_delayed_node_to_data_ref(ref);
744 	full_ref->parent = parent;
745 	full_ref->root = ref_root;
746 	if (parent)
747 		ref->type = BTRFS_SHARED_DATA_REF_KEY;
748 	else
749 		ref->type = BTRFS_EXTENT_DATA_REF_KEY;
750 
751 	full_ref->objectid = owner;
752 	full_ref->offset = offset;
753 
754 	trace_add_delayed_data_ref(fs_info, ref, full_ref, action);
755 
756 	ret = insert_delayed_ref(trans, delayed_refs, head_ref, ref);
757 	if (ret > 0)
758 		kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
759 }
760 
761 /*
762  * add a delayed tree ref.  This does all of the accounting required
763  * to make sure the delayed ref is eventually processed before this
764  * transaction commits.
765  */
766 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
767 			       struct btrfs_trans_handle *trans,
768 			       u64 bytenr, u64 num_bytes, u64 parent,
769 			       u64 ref_root,  int level, int action,
770 			       struct btrfs_delayed_extent_op *extent_op,
771 			       int *old_ref_mod, int *new_ref_mod)
772 {
773 	struct btrfs_delayed_tree_ref *ref;
774 	struct btrfs_delayed_ref_head *head_ref;
775 	struct btrfs_delayed_ref_root *delayed_refs;
776 	struct btrfs_qgroup_extent_record *record = NULL;
777 	int qrecord_inserted;
778 	int is_system = (ref_root == BTRFS_CHUNK_TREE_OBJECTID);
779 
780 	BUG_ON(extent_op && extent_op->is_data);
781 	ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
782 	if (!ref)
783 		return -ENOMEM;
784 
785 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
786 	if (!head_ref)
787 		goto free_ref;
788 
789 	if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
790 	    is_fstree(ref_root)) {
791 		record = kmalloc(sizeof(*record), GFP_NOFS);
792 		if (!record)
793 			goto free_head_ref;
794 	}
795 
796 	head_ref->extent_op = extent_op;
797 
798 	delayed_refs = &trans->transaction->delayed_refs;
799 	spin_lock(&delayed_refs->lock);
800 
801 	/*
802 	 * insert both the head node and the new ref without dropping
803 	 * the spin lock
804 	 */
805 	head_ref = add_delayed_ref_head(fs_info, trans, head_ref, record,
806 					bytenr, num_bytes, 0, 0, action, 0,
807 					is_system, &qrecord_inserted,
808 					old_ref_mod, new_ref_mod);
809 
810 	add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
811 			     num_bytes, parent, ref_root, level, action);
812 	spin_unlock(&delayed_refs->lock);
813 
814 	if (qrecord_inserted)
815 		btrfs_qgroup_trace_extent_post(fs_info, record);
816 
817 	return 0;
818 
819 free_head_ref:
820 	kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
821 free_ref:
822 	kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
823 
824 	return -ENOMEM;
825 }
826 
827 /*
828  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
829  */
830 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
831 			       struct btrfs_trans_handle *trans,
832 			       u64 bytenr, u64 num_bytes,
833 			       u64 parent, u64 ref_root,
834 			       u64 owner, u64 offset, u64 reserved, int action,
835 			       int *old_ref_mod, int *new_ref_mod)
836 {
837 	struct btrfs_delayed_data_ref *ref;
838 	struct btrfs_delayed_ref_head *head_ref;
839 	struct btrfs_delayed_ref_root *delayed_refs;
840 	struct btrfs_qgroup_extent_record *record = NULL;
841 	int qrecord_inserted;
842 
843 	ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
844 	if (!ref)
845 		return -ENOMEM;
846 
847 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
848 	if (!head_ref) {
849 		kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
850 		return -ENOMEM;
851 	}
852 
853 	if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
854 	    is_fstree(ref_root)) {
855 		record = kmalloc(sizeof(*record), GFP_NOFS);
856 		if (!record) {
857 			kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
858 			kmem_cache_free(btrfs_delayed_ref_head_cachep,
859 					head_ref);
860 			return -ENOMEM;
861 		}
862 	}
863 
864 	head_ref->extent_op = NULL;
865 
866 	delayed_refs = &trans->transaction->delayed_refs;
867 	spin_lock(&delayed_refs->lock);
868 
869 	/*
870 	 * insert both the head node and the new ref without dropping
871 	 * the spin lock
872 	 */
873 	head_ref = add_delayed_ref_head(fs_info, trans, head_ref, record,
874 					bytenr, num_bytes, ref_root, reserved,
875 					action, 1, 0, &qrecord_inserted,
876 					old_ref_mod, new_ref_mod);
877 
878 	add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
879 				   num_bytes, parent, ref_root, owner, offset,
880 				   action);
881 	spin_unlock(&delayed_refs->lock);
882 
883 	if (qrecord_inserted)
884 		return btrfs_qgroup_trace_extent_post(fs_info, record);
885 	return 0;
886 }
887 
888 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
889 				struct btrfs_trans_handle *trans,
890 				u64 bytenr, u64 num_bytes,
891 				struct btrfs_delayed_extent_op *extent_op)
892 {
893 	struct btrfs_delayed_ref_head *head_ref;
894 	struct btrfs_delayed_ref_root *delayed_refs;
895 
896 	head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
897 	if (!head_ref)
898 		return -ENOMEM;
899 
900 	head_ref->extent_op = extent_op;
901 
902 	delayed_refs = &trans->transaction->delayed_refs;
903 	spin_lock(&delayed_refs->lock);
904 
905 	/*
906 	 * extent_ops just modify the flags of an extent and they don't result
907 	 * in ref count changes, hence it's safe to pass false/0 for is_system
908 	 * argument
909 	 */
910 	add_delayed_ref_head(fs_info, trans, head_ref, NULL, bytenr,
911 			     num_bytes, 0, 0, BTRFS_UPDATE_DELAYED_HEAD,
912 			     extent_op->is_data, 0, NULL, NULL, NULL);
913 
914 	spin_unlock(&delayed_refs->lock);
915 	return 0;
916 }
917 
918 /*
919  * this does a simple search for the head node for a given extent.
920  * It must be called with the delayed ref spinlock held, and it returns
921  * the head node if any where found, or NULL if not.
922  */
923 struct btrfs_delayed_ref_head *
924 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
925 {
926 	return find_ref_head(&delayed_refs->href_root, bytenr, 0);
927 }
928 
929 void __cold btrfs_delayed_ref_exit(void)
930 {
931 	kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
932 	kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
933 	kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
934 	kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
935 }
936 
937 int __init btrfs_delayed_ref_init(void)
938 {
939 	btrfs_delayed_ref_head_cachep = kmem_cache_create(
940 				"btrfs_delayed_ref_head",
941 				sizeof(struct btrfs_delayed_ref_head), 0,
942 				SLAB_MEM_SPREAD, NULL);
943 	if (!btrfs_delayed_ref_head_cachep)
944 		goto fail;
945 
946 	btrfs_delayed_tree_ref_cachep = kmem_cache_create(
947 				"btrfs_delayed_tree_ref",
948 				sizeof(struct btrfs_delayed_tree_ref), 0,
949 				SLAB_MEM_SPREAD, NULL);
950 	if (!btrfs_delayed_tree_ref_cachep)
951 		goto fail;
952 
953 	btrfs_delayed_data_ref_cachep = kmem_cache_create(
954 				"btrfs_delayed_data_ref",
955 				sizeof(struct btrfs_delayed_data_ref), 0,
956 				SLAB_MEM_SPREAD, NULL);
957 	if (!btrfs_delayed_data_ref_cachep)
958 		goto fail;
959 
960 	btrfs_delayed_extent_op_cachep = kmem_cache_create(
961 				"btrfs_delayed_extent_op",
962 				sizeof(struct btrfs_delayed_extent_op), 0,
963 				SLAB_MEM_SPREAD, NULL);
964 	if (!btrfs_delayed_extent_op_cachep)
965 		goto fail;
966 
967 	return 0;
968 fail:
969 	btrfs_delayed_ref_exit();
970 	return -ENOMEM;
971 }
972